Michele Ong
Generate SRT files with Google Cloud Services

Generate SRT files with Google Cloud Services

tl;dr: I need to generate my own SRT files for my videos because Facebook Creator Studio makes me angry.

Long: Backstory.


More details for installation and usage in the script below.


Default GCS only works for videos no longer than 60 seconds, and the script will only send the first 60 seconds of video to accommodate this. Modifications are required to use GCS for longer videos[1].

I find the video transcription model more accurate than the default, but video is a premium model if that matters to you. I still need to make corrections, but far fewer than Facebook’s auto-captioning feature.

For bin length, I’ve set it to a default of 3 seconds with a line break of 37 characters[2].

00:00:05,100 --> 00:00:07,500
Big Bang Theory thank you. So
especially in that
Check out STEAM Powered (https://steampoweredshow.com/) where I have conversations
with women in STEAM to learn a bit about what they do and who they are.
* Install `ffmpeg <https://www.ffmpeg.org/>`_
* Setup virtual env `<https://cloud.google.com/python/docs/setup>`_
* `pip install google-cloud-speech`
* `pip install pydub`
* `pip install srt`
You will also need API Credentials with Cloud Speech-to-Text API enabled.
Instructions here: `<https://cloud.google.com/speech-to-text/docs/quickstart-client-libraries>`_
**Disclaimer:** Be aware that this is technically not a free service. Check Google's pricing to confirm whether this service
is suitable for your requirements.
* This has been tested with .mov, and that's it.
* Accuracy determined by the quality of the source video, and the service.
* Video length cannot exceed 60 seconds with this Google service. The `ffmpeg` command will only send the first 60 seconds.
usage: makesrt.py [-h] [-b BIN] -c CREDENTIALS [-l [LANG]] [-m [{command_and_search,phone_call,video,default}]] [-r BREAK] [filename]
positional arguments:
filename video filename
optional arguments:
-h, --help show this help message and exit
-b BIN, --bin BIN caption length, defaults to 3 seconds
Google credentials JSON
-l [LANG], --lang [LANG]
language code, defaults to en-US. See https://cloud.google.com/speech-to-text/docs/languages
-m [{command_and_search,phone_call,video,default}], --model [{command_and_search,phone_call,video,default}]
transcription model, defaults to default. Note: video is a premium model.
-r BREAK, --break BREAK
line break length, defaults to 37 characters
Cobbled together from -
* `autosub <https://github.com/agermanidis/autosub/>`_
* `<https://github.com/darshan-majithiya/Generate-SRT-File-using-Google-Cloud-s-Speech-to-Text-API>`_
import argparse, datetime, io, os, subprocess, sys, tempfile, textwrap
from google.cloud import speech
from pydub.utils import mediainfo
import srt
def build_subs(response):
transcript= []
for result in response.results:
for alternative in result.alternatives:
if not alternative:
""" Initialise timing """
if alternative.words[0].start_time.seconds:
start_time = alternative.words[0].start_time
start_time = datetime.timedelta(seconds=0, microseconds=0)
end_time = alternative.words[-1].end_time
subtitle = ''
previous_word = None
for word in alternative.words:
if word.end_time < start_time + datetime.timedelta(seconds=args.bin):
subtitle = subtitle + " " + word.word
elif previous_word:
subtitle = textwrap.dedent(subtitle)
subtitle = '\n'.join(l for line in subtitle.splitlines()
for l in textwrap.wrap(line, width=37)
srt.Subtitle(len(transcript) + 1, start_time, previous_word.end_time, subtitle)
start_time = word.start_time
end_time = word.end_time + datetime.timedelta(seconds=args.bin)
subtitle = word.word
previous_word = word
except IndexError:
# Last one
srt.Subtitle(len(transcript) + 1, start_time, previous_word.end_time, subtitle)
except IndexError:
return srt.compose(transcript)
parser = argparse.ArgumentParser()
parser.add_argument('-b', '--bin', help='caption length, defaults to 3 seconds', default=3)
parser.add_argument('-c', '--credentials', help='Google credentials JSON', required=True)
parser.add_argument('-l', '--lang', help='language code, defaults to en-US. See https://cloud.google.com/speech-to-text/docs/languages', nargs='?', default='en-US')
help='transcription model, defaults to default. Note: video is a premium model.',
choices=['command_and_search', 'phone_call', 'video', 'default'],
parser.add_argument('-r', '--linebreak', help='line break length, defaults to 37 characters', default=37)
parser.add_argument('filename', help='video filename', nargs='?', type=str)
args = parser.parse_args()
os.environ["GOOGLE_APPLICATION_CREDENTIALS"] = args.credentials
file_name = args.filename
""" Generate the temporary audio file """
video_data = mediainfo(file_name)
channels = video_data["channels"]
bit_rate = video_data["bit_rate"]
sample_rate = video_data["sample_rate"]
temp_file = tempfile.NamedTemporaryFile(suffix='.flac', delete=False)
command = f"ffmpeg -i \"{file_name}\" -ss 0 -t 60 -b:a {bit_rate} -ac {channels} -ar {sample_rate} -vn \"{temp_file.name}\""
subprocess.call(command, shell=True)
""" Prepare the audio """
with io.open(temp_file.name, "rb") as audio_file:
content = audio_file.read()
audio = speech.RecognitionAudio(content=content)
config = speech.RecognitionConfig(
language_code = args.lang,
sample_rate_hertz = int(sample_rate),
encoding = "FLAC",
audio_channel_count = int(channels),
enable_word_time_offsets = True,
model = args.model,
enable_automatic_punctuation = True
client = speech.SpeechClient()
response = client.recognize(request={"config": config, "audio": audio})
""" Build and output SRT """
subs = build_subs(response)
basename, ext = os.path.splitext(file_name)
lang = args.lang.replace('-', '_')
with open("{basename}.{lang}.srt".format(basename=basename, lang=lang), "w") as f:
view raw makesrt.py hosted with ❤ by GitHub

[1] See https://towardsdatascience.com/how-to-use-google-speech-to-text-api-to-transcribe-long-audio-files-1c886f4eb3e9

[2] 3 seconds seems to come up a bit, and the first subtitle style-guide I found is from the BBC and I saw 37 characters on there because of teletext. For online it’s based on percentage of screen-width and at that point I realised close enough is good enough. Got better numbers? Let me know.

Published December 04, 2020
STEAM PoweredcaptionsGCSgist

© 2022 Michele Ong.